Re: [linus:master] [selftests/cgroup] 954bacce36: kernel-selftests.cgroup.test_cpu.test_cpucg_max.fail

[Michal Koutný]

Hello.

On Thu, Aug 07, 2025 at 01:52:31PM +0800, kernel test robot <oliver.sang@xxxxxxxxx> wrote:
> dfe25fbaedfc2a07 954bacce36d976fe472090b5598
> ---------------- ---------------------------
>        fail:runs  %reproduction    fail:runs
>            |             |             |
>            :6           100%          6:6     kernel-selftests.cgroup.test_cpu.test_cpucg_max.fail
>            :6           100%          6:6     kernel-selftests.cgroup.test_cpu.test_cpucg_max_nested.fail
> 
> 
> not sure if there are any necessary env setting? thanks

The selftest commit essentially changed the tolerance margin from
ridiculously large to something that looked statistically appropriate
[1].
However, when I run the test (30x) on the 954bacce36 I get:

quantile([D1 D2 D8])  # 1 2 and 8 vCPUs respectively
ans =

   1.3086e+04   1.1559e+04   1.1177e+04 # min
   1.5109e+04   1.2936e+04   1.2989e+04
   1.5791e+04   1.3938e+04   1.3788e+04 # median
   1.6159e+04   1.5385e+04   1.4980e+04
   1.8757e+04   1.8699e+04   1.9494e+04 # max

I obtain similar values also on v6.15 (the kernel + 954bacce36
selftest). So it's not anything in throtlling implementation affecting
this.

The tests above were with HZ=250, for HZ=1000, I get slightly smaller
results with D2:
   1.1753e+04 # min
   1.2634e+04
   1.3208e+04 # median
   1.4010e+04
   1.6937e+04 # max

But still nowhere the 20% margin (i.e. values_close(...10%)), these
values would demand up to 100% (values_close(..., 50%)). Or add a bias
derived from sched_cfs_bandwidth_slice_us or increase the tested quota
from 1% to 5%, that'd be an improvement:

   48882 # min
   52450
   52941 # median
   54284 # 75th percentile
   73186 # max (limit would be 60000)

I'm not sure how big overrun we want to accept as a pass.

Michal

[1] lore.kernel.org/r/20250701-kselftest-cgroup-fix-cpu-max-v1-2-049507ad6832@xxxxxxxx

Attachment: signature.asc
Description: PGP signature